Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.comfy.org/llms.txt

Use this file to discover all available pages before exploring further.

Uni-1 is Luma’s first unified model designed for both image creation and precision editing. Unlike most image models, Uni-1 isn’t a diffusion model — it’s a decoder-only autoregressive transformer that treats text and images as a single interleaved sequence, jointly modeling time, space, and logic in one architecture. The result: a model that reasons about your prompt before generating — decomposing instructions, resolving constraints, and planning composition like a frontier LLM. It thinks first, then draws. Beyond raw output quality, what makes it powerful is control: you can generate entirely new images or surgically modify existing ones, while guiding the system with references, seeds, and structured prompts. This guide gives you a working mental model and practical workflows, from your first image to advanced multi-reference setups.

​
What makes Uni-1 different

  • Two clear modes: Create Image (generate something new) and Modify Image (edit something existing)
  • Up to 9 reference images, each with a defined role
  • Strong control over niche and specific visual styles
  • Seed support for reproducibility and controlled iteration
  • Nine aspect ratios, from ultra-tall to ultra-wide
  • Text rendering that’s actually readable
  • Web search grounding for real-world context
  • Multilingual prompts
  • Multi-panel output with temporal consistency
If you understand the modes and how to control references, you understand Uni-1.

​
Strengths

Uni-1 excels across a wide range of tasks:
  • Photorealism with material accuracy
  • Illustration & stylized art with strong aesthetic control
  • Old photo restoration and vintage reproduction
  • Surreal and conceptual compositions
  • Text rendering — readable text inside images, great for infographics and posters
  • Image editing and multi-turn refinement
  • Reference-guided generation with identity preservation
  • Multi-panel output — consistent characters/scenes across multiple frames

​
The core distinction: Create vs Modify

Everything in Uni-1 starts with one question: Am I creating something new, or changing something that already exists?
ModeWhat it doesWhen to use
Create ImageProduces a brand-new composition. Can be inspired by references.”Create a new scene in the style of this photo”
Modify ImageEdits a specific input image. Preserves composition and structure unless told otherwise.”Make this photo look like nighttime”
When in doubt:
  • If the output should look like a version of your input, use Modify
  • If it should feel inspired but new, use Create

​
Getting started

You can use Uni-1 in two ways:
  • Comfy Cloud: Preview the template and run it directly in your browser.
  • Desktop: Update ComfyUI to the latest version, find the Luma UNI-1 Image node via the Node Library, and connect it to your workflow.

​
Image Create workflow

Run Image Create on Cloud

Try the Image Create workflow instantly on Comfy Cloud.

Download Image Create Workflow

Download JSON or search “Luma UNI-1 Image Create” in Template Library

​
Image Edit workflow

Run Image Edit on Cloud

Try the Image Edit workflow instantly on Comfy Cloud.

Download Image Edit Workflow

Download JSON or search “Luma UNI-1 Image Edit” in Template Library
The workflow is simple: prompt → evaluate → refine. Leave the seed blank while exploring. Once you find something strong, lock the seed and iterate from there.

​
Core parameters

ParameterDescription
PromptYour primary control. Up to 6,000 characters. Be precise.
Aspect ratioControls framing, not quality. Choose based on use case.
SeedSame seed + same prompt → same result. Same seed + changed prompt → controlled variation. No seed → exploration.
Reference images (Create)Up to 9 images to guide different aspects.
Source image (Modify)The image you are editing. Dimensions are preserved automatically.

​
Working with reference images

References only work if you tell the model what they are for. Use this structure: 
Use IMAGE1 ([description]) as a [ROLE] reference.
Possible roles: Style, Character, Composition, Color palette, Lighting, Texture, Mood. Without roles, the model guesses, and guesses are unreliable.

​
Create mode examples

Style reference 
Use IMAGE1 (impressionist oil painting with loose brushwork and warm
sunset tones) as a STYLE reference. Create a portrait of a woman in her
30s applying IMAGE1's color palette and painterly texture. The subject
should be new.
Character reference 
Use IMAGE1 (woman with short copper-red hair, freckles) as a CHARACTER
reference. Preserve her features. Generate a new scene: she's sitting
in a softly lit café.
Multi-reference 
Use IMAGE1 as a COLOR PALETTE reference, IMAGE2 as LIGHTING, IMAGE3 as
COMPOSITION. Create a lone figure walking through a rain-slicked street
at night.

​
Modify mode examples

In Modify mode, clarity is everything. Always specify what to change and what must stay untouched. 
Change the time of day to golden hour. Update sky, light direction,
shadows, and color temperature. Keep all subjects and composition
unchanged.

Make the rain heavier and add reflections on the ground. Keep all other
elements unchanged.

​
Prompting guidelines

Recommended lengths:
  • Text-to-image → 80–250 words
  • Reference-guided → 100–300 words
  • Modify → 30–100 words
Avoid vague terms (“beautiful”, “amazing”), redundant phrasing, and conflicting instructions. Instead, use specific, named aesthetics:
  • Golden hour, 85mm lens, shallow depth of field
  • 1970s Italian giallo film poster, high-contrast color blocking
Precision beats generality.

​
Aspect ratios

Choose based on where the image will live:
RatioUse case
1:1Social posts
9:16Vertical video
16:9Widescreen
3:2 / 2:3Photography
2:1 / 3:1Cinematic / panoramic
1:2 / 1:3Ultra-tall
In Modify mode, aspect ratio is locked to the source image.

​
Seeds: control and reproducibility

Seeds turn experimentation into systems.
  • Fixed seed → consistency
  • No seed → exploration
Workflow:
  1. Explore with no seed
  2. Find a strong result
  3. Lock the seed
  4. Change one variable at a time
Save your prompt and seed together. That’s your reusable recipe.

​
Advanced techniques

​
Character consistency

  1. Generate a clean, front-facing reference image
  2. Reuse it as IMAGE1 (CHARACTER) in every scene
  3. Keep the label identical across prompts

​
Multi-reference architecture

Assign one role per image:
  • IMAGE1 → character
  • IMAGE2 → style
  • IMAGE3 → lighting
  • IMAGE4 → environment
End the prompt with: 
Treat each reference as having authority over its assigned layer only.

​
Create → Modify chain

Use Create to explore compositions, then Modify to refine details. This is one of the most powerful workflows.

​
Iterative refinement

  1. Explore (no seed)
  2. Lock seed
  3. Change one variable per generation
  4. Document results
It feels slower, but it’s actually faster, because you always know what changed.

​
Troubleshooting

ProblemFix
References ignoredLabel each one clearly
Modify changes too muchExplicitly state what must stay unchanged
Inconsistent outputsLock the seed
Prompt partially ignoredRemove conflicts or split into steps
Output looks like the reference imageYou may be in Modify mode
Character inconsistencyReuse a canonical reference image

​
Quick reference

Create Image
  • New compositions
  • Text + up to 9 references
  • Descriptive prompts
Modify Image
  • Edits existing images
  • Source image + references
  • Direct, surgical prompts

​
The golden rules

  1. Label every reference
  2. In Modify mode, always state what should not change
  3. Change one variable at a time when refining
  4. Save prompt + seed for reproducibility
  5. Create = new scenes, Modify = edits

​
Key takeaway

Mastering Uni-1 comes down to three things: choose the right mode → control your references → iterate with intention. Once those are in place, you’re no longer guessing, you’re directing.